A detailed empirical comparison of token consumption between Claude Code and OpenCode, measured at the API boundary using a logging proxy. Claude Code's baseline overhead is ~33,000 tokens per request vs OpenCode's ~7,000, driven by 27 tool schemas and injected scaffolding. Cache instability in Claude Code causes 5.9x–54x more cache writes than OpenCode on identical tasks. Real-world configurations (instruction files, MCP servers, subagents) can push first-request overhead to 75,000–90,000 tokens. Subagent fan-out multiplies costs dramatically: a 121k-token direct task ballooned to 513k tokens with two subagents. One counterpoint: Claude Code's parallel tool batching can make multi-step tasks cheaper than OpenCode's serial one-call-per-turn approach. The post includes methodology details, caveats, and instructions for reproducing the measurement rig.
